Hello readers!

This week, we’re looking at what we can learn from weed-whacking robots, what SGP.32 means for automotive connectivity, and more!

A weed-picking robot has to tell a weed apart from a crop, in a field that looks nothing like the field two counties over, using a camera that can't phone home for help. That's the challenge underneath the autonomous rovers NVIDIA highlighted during National Robotics Week, which use vision AI to spot and remove weeds without herbicide. The rovers themselves are almost beside the point. What makes the system work is a pairing showing up across outdoor robotics right now — synthetic data to handle the variety of real-world conditions, and on-device inference to handle the fact that most of those conditions have no reliable connection back to a server.

Start with the variety problem, since it's the one people underestimate. A vision model trained on soybean fields in one region can fail on lettuce rows in another, not because the model is bad but because the training distribution never saw that combination of soil, weed species, and light. Collecting enough real footage to cover every crop and geography is slow and often impractical. The workaround is to generate the missing scenarios rather than film them, sampling world models to produce synthetic variations. Tonic.ai's Adam Kamor described this same logic, framing synthetic data as a stand-in wherever real examples are too scarce, costly, or sensitive to gather at scale. Weeds aren't sensitive, but they're scarce in the specific combinations any one model needs to see.

Solving the data problem still leaves the deployment problem, and it's arguably the harder one. A model that performs well in a lab means little if it can't run where the robot actually is. Cloud-dependent inference adds a few hundred milliseconds of round-trip latency and stops working the moment a connection drops — tolerable for a recommendation engine, disqualifying for a machine making physical decisions in a field. IoT For All's breakdown of Edge AI and Tiny ML in robotics lays out why this pushes intelligence onto the device itself: local processing removes the latency, connectivity dependency, and bandwidth strain in one move, at the cost of needing a vision model compact enough for modest, often solar-powered hardware.

That compression work is easy to miss because it's invisible in a demo video, but it's where a lot of the real engineering effort goes — and it's what lets these systems do something older IoT agriculture tools couldn't. Soil moisture, leaf wetness, and weather sensors have been standard in precision farming for years, but none of them could see. They could flag that conditions favored disease; they couldn't point at the plant.

The pattern extends well past agriculture. Any system meant to operate somewhere unpredictable and disconnected — a field, a construction site, a remote pipeline — hits the same two constraints before it hits anything resembling a hard AI problem: where does representative training data come from when the real world won't hold still, and how does the model get small enough to live on the device instead of a server rack. Get those right, and the computer vision part is often the easy half.

Remote SIM provisioning has become a strategic enabler for large-scale IoT and automotive deployments. As connected devices grow in number, longevity, and geographic reach, traditional SIM management approaches are increasingly unable to keep pace. With SGP.32, the GSMA has introduced a specification purpose-built for the operational, technical, and commercial realities of IoT and automotive connectivity.

Read more

When you add AI to your app, the data profile changes overnight. Every prompt, response, and user interaction becomes a timestamped event. That's not your app's usual row count.

Vanilla Postgres handles it until it doesn't. Query times creep up. Dashboard refreshes slow down. You start reaching for a second database or a data pipeline to offload the load.

TimescaleDB extends Postgres for exactly this. It doesn't replace what's working. It makes Postgres stay fast as AI-generated data piles up.

Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage 10x. Continuous aggregates keep your dashboards live without re-querying everything. No pipeline. No second database. No migration.

Same Postgres. Same SQL. Just built to handle what AI features actually generate.

Get $1000 Credit To Start

In this episode of the IoT For All Podcast, Martin Nord, Chief Technology and Product Officer at Com4, joins Ryan Chacon to discuss IoT in 2026 and what enterprises need to consider. The conversation covers what enterprise IoT buyers actually need, the crowded IoT connectivity market, the 2G/3G sunset, how roaming in IoT is broken in subtle ways, the nature of IoT connectivity failures, prioritizing financial stability in IoT partners, satellite IoT, and the future of IoT connectivity.

YouTube • Spotify • Apple • YT Music • Amazon

Interested in becoming an IoT For All Partner? Reach out here!